Road segmentation system and method based on full-convolution neural network and condition random field

Abstract

The invention discloses a road segmentation system and method based on a full-convolution neural network and a condition random field. The method comprises the following steps: constructing a full-convolution network based on the VGG_16 deep convolution network by regarding the road segmentation as a dichotomy problem by using good feature representation capacity of the deep neural network for realizing the end-to-end road surface and background classification of the road image; and performing smooth optimization on a rough edge obtained through dichotomy by using a full-connection condition random field by utilizing the characteristic that the image fine segmentation can be realized by using the full-connection condition random field. Through the method disclosed by the invention, the segmentation accuracy rate of 98.13% and the segmentation speed of processing one image per 0.84s are acquired, and an efficient solution is provided for the traffic scene image road segmentation.

Description

technical field [0001] The invention belongs to the field of computer vision, and in particular relates to a road segmentation system and method based on a fully convolutional neural network and a conditional random field. Background technique [0002] Road segmentation has always been an important topic in the field of intelligent driving research. Traditional road segmentation methods are mostly based on the research of inherent image attributes such as color, texture, edge, and road geometry. Color-based segmentation algorithms include robust Gauss method, K-means clustering, nearest neighbor method, etc. Such algorithms are often sensitive to interference such as road shadows and ponding, and the segmentation performance is easily attenuated. Catmull-Rom spline model and B-Snake model, this kind of method requires the road edge to be clear in order to facilitate the selection of control points, but this kind of method relies too much on the control points, the modeling...

AI Technical Summary

Problems solved by technology

However, the traditional convolutional neural network uses a large receptive field, resulting in rough edges in image segmentation results. At the same time, due to the use of maximum pooling, each pooling layer only extracts the most obvious features, which further aggravates the problem of rough edges in segmentation.

The conditional random field was proposed by Lafferty et al. for the labeling problem. It is a discriminative probabilistic undirected graph learning model, but it only has advantages for the combination of similar features or similar features, and it usually has long-distance dependence on the distribution of observation data. It has good labeling performance only under the condition, and cannot be well applied to other situations

Images